chore: stash UI improvements before test images

.gitignore

@@ -0,0 +1 @@
+.agents/

scripts/modeling_vit_classifier.py

@@ -1,3 +1,9 @@
+# DEPRECATED — July 2026
+# This wrapper is kept for backwards compatibility and archived in scripts/.
+# Standard inference should use the HuggingFace path:
+#   from transformers import ViTForImageClassification, ViTImageProcessor
+#   model = ViTForImageClassification.from_pretrained("buildborderless/CommunityForensics-DeepfakeDet-ViT")
+#   processor = ViTImageProcessor.from_pretrained("buildborderless/CommunityForensics-DeepfakeDet-ViT")
 from transformers import PreTrainedModel
 import torch
 import torch.nn as nn

test_app/app.py

@@ -68,8 +68,8 @@ def _onnx_preprocess(image):
 
 # ── Tab 1: PyTorch vs ONNX comparison ─────────────────────────────────
 
-def compare(image, onnx_variant):
-    if image is None:
+def compare(image, onnx_variants):
+    if image is None or not onnx_variants:
         return (None, None)
 
     t0 = time.perf_counter()
@@ -89,24 +89,25 @@ def compare(image, onnx_variant):
         "time_ms": round(pt_ms, 1),
     }
 
-    t0 = time.perf_counter()
-    session = _get_onnx(onnx_variant)
-    arr = _onnx_preprocess(image)
-    use_f16 = session.get_inputs()[0].type == "tensor(float16)"
-    inp = arr.astype(np.float16) if use_f16 else arr
-    onx_logits = session.run(None, {"pixel_values": inp})[0][0].astype(np.float32)
-    onx_probs = np.exp(onx_logits - onx_logits.max()) / np.exp(onx_logits - onx_logits.max()).sum()
-    onx_pred = "real" if np.argmax(onx_probs) == 0 else "fake"
-    onx_ms = (time.perf_counter() - t0) * 1000
-
-    onx = {
-        "device": onnx_device,
-        "variant": onnx_variant,
-        "prediction": onx_pred,
-        "real": round(float(onx_probs[0]), 4),
-        "fake": round(float(onx_probs[1]), 4),
-        "time_ms": round(onx_ms, 1),
-    }
+    onx = {}
+    for variant in onnx_variants:
+        session = _get_onnx(variant)
+        arr = _onnx_preprocess(image)
+        use_f16 = session.get_inputs()[0].type == "tensor(float16)"
+        t0 = time.perf_counter()
+        inp = arr.astype(np.float16) if use_f16 else arr
+        onx_logits = session.run(None, {"pixel_values": inp})[0][0].astype(np.float32)
+        onx_probs = np.exp(onx_logits - onx_logits.max()) / np.exp(onx_logits - onx_logits.max()).sum()
+        onx_pred = "real" if np.argmax(onx_probs) == 0 else "fake"
+        onx_ms = (time.perf_counter() - t0) * 1000
+
+        onx[variant] = {
+            "device": onnx_device,
+            "prediction": onx_pred,
+            "real": round(float(onx_probs[0]), 4),
+            "fake": round(float(onx_probs[1]), 4),
+            "time_ms": round(onx_ms, 1),
+        }
 
     return (pt, onx)
 
@@ -199,15 +200,16 @@ with gr.Blocks(title="DeepfakeDet-ViT") as demo:
     with gr.Tabs():
         with gr.TabItem("Compare"):
             with gr.Row():
-                img_in = gr.Image(type="pil", label="Upload Image", scale=1)
-            variant = gr.Dropdown(
-                choices=list(ONNX_VARIANTS.keys()),
-                value=list(ONNX_VARIANTS.keys())[0],
-                label="ONNX Variant",
-            )
-            with gr.Row():
-                pt_out = gr.JSON(label="PyTorch", scale=1)
-                onx_out = gr.JSON(label="ONNX", scale=1)
+                with gr.Column(scale=1):
+                    img_in = gr.Image(type="pil", label="Upload Image")
+                with gr.Column(scale=1):
+                    variant = gr.CheckboxGroup(
+                        choices=list(ONNX_VARIANTS.keys()),
+                        value=[list(ONNX_VARIANTS.keys())[2]],
+                        label="ONNX Variants (select 1+)",
+                    )
+                    pt_out = gr.JSON(label="PyTorch")
+                    onx_out = gr.JSON(label="ONNX")
             img_in.change(fn=compare, inputs=[img_in, variant], outputs=[pt_out, onx_out])
             variant.change(fn=compare, inputs=[img_in, variant], outputs=[pt_out, onx_out])